Merge pull request #14994 from nextcloud/backport/14923/stable34

[stable34] docs: expand PHP-FPM tuning section with real guidance

Author: skjnldsv

admin_manual/installation/server_tuning.rst

@@ -139,12 +139,135 @@ HTTP/2 has `huge speed improvements <https://www.troyhunt.com/i-wanna-go-fast-ht
 Tune PHP-FPM
 ------------
 
-The default configuration of PHP-FPM is extremely conservative. You might notice
-excessive load times on the web interface or even sync issues. Each simultaneous
-request is handled by a separate PHP-FPM process, so even on a small installation you
-should allow more processes to run in parallel to handle requests.
+PHP-FPM is required for Nginx setups and is widely used with Apache as well. Its default
+configuration is extremely conservative: the default pool has ``pm.max_children = 5``, which
+limits Nextcloud to five simultaneous PHP requests and is a common cause of gateway timeouts,
+slow page loads, and sync client errors under any real load.
 
-`This link <https://spot13.com/pmcalculator/>`_ can help you calculate the optimal values for your system.
+Process manager modes
+^^^^^^^^^^^^^^^^^^^^^
+
+The ``pm`` directive controls how PHP-FPM manages its worker processes:
+
+``dynamic``
+  Keeps between ``pm.min_spare_servers`` and ``pm.max_spare_servers`` idle workers alive,
+  up to a ceiling of ``pm.max_children``. Good default for most Nextcloud installations:
+  balances RAM efficiency with burst capacity. Set ``pm.min_spare_servers`` high enough
+  that sync-client poll bursts do not stall waiting for new processes to spawn.
+
+``static``
+  Always keeps exactly ``pm.max_children`` processes running. Highest memory use,
+  lowest latency. Use on dedicated servers with predictable load. Always set
+  ``pm.max_requests`` to recycle workers and prevent memory leaks.
+
+``ondemand``
+  Spawns a worker only when a request arrives; kills idle workers after
+  ``pm.process_idle_timeout`` (default ``10s``). Lowest memory use but adds
+  cold-start latency on every burst. Not recommended for Nextcloud: desktop and
+  mobile clients poll every 30 seconds, repeatedly triggering cold starts.
+
+Key parameters
+^^^^^^^^^^^^^^
+
+``pm.max_children``
+  Maximum (or fixed, under ``static``) number of simultaneous worker processes.
+  This is the most important value to tune. If all workers are busy, new requests
+  queue up; a full queue produces 502/504 errors.
+
+  Estimate it from available RAM::
+
+    pm.max_children = floor(available_RAM_for_PHP / average_worker_RSS)
+
+  Measure the average RSS of a running pool::
+
+    ps --no-headers -o rss -C php-fpm | awk '{sum+=$1; count++} END {if (count>0) print sum/count/1024 " MB"; else print "No php-fpm processes found"}'
+
+  A typical Nextcloud worker uses **50–100 MB** (more if Imagick or LDAP is loaded).
+  Leave headroom for the OS, web server, database, and cache. Setting ``pm.max_children``
+  too high causes swapping, which is worse than queuing.
+
+``pm.start_servers`` *(dynamic only)*
+  Workers started at FPM boot. Defaults to
+  ``(pm.min_spare_servers + pm.max_spare_servers) / 2`` if not set.
+
+``pm.min_spare_servers`` / ``pm.max_spare_servers`` *(dynamic only)*
+  Range of idle workers kept warm. For Nextcloud, keep ``pm.min_spare_servers``
+  high enough to absorb a sync-client burst without spawning new processes::
+
+    pm.min_spare_servers = 4    # adjust upward for many connected clients
+    pm.max_spare_servers = 16
+
+``pm.max_requests``
+  Recycle a worker after this many requests. ``0`` means never recycle.
+  Setting a value of ``500``–``1000`` guards against slow memory growth from
+  leaky extensions (Imagick, LDAP, SAML XML parsers). Essential under ``static``
+  mode.
+
+``pm.process_idle_timeout`` *(ondemand only)*
+  How long an idle worker lives before being killed. Default: ``10s``.
+
+Example configuration
+^^^^^^^^^^^^^^^^^^^^^
+
+A starting point for ``dynamic`` mode on a server with 2 GB of RAM dedicated to PHP
+(adjust ``pm.max_children`` based on your measured worker RSS):
+
+.. code-block:: ini
+
+   pm = dynamic
+   pm.max_children = 30
+   pm.start_servers = 8
+   pm.min_spare_servers = 4
+   pm.max_spare_servers = 16
+   pm.max_requests = 500
+
+Use the `PHP-FPM process calculator <https://spot13.com/pmcalculator/>`_ as a
+cross-check for your values.
+
+Slow log
+^^^^^^^^
+
+Enable the slow log to identify PHP scripts that are taking too long:
+
+.. code-block:: ini
+
+   slowlog = /var/log/php-fpm-slow.log
+   request_slowlog_timeout = 5s
+
+Each entry records the full PHP backtrace of the slow request. This is the
+fastest way to find the root cause of gateway timeouts and sluggish pages.
+
+Troubleshooting
+^^^^^^^^^^^^^^^
+
+**502 Bad Gateway**
+  All ``pm.max_children`` workers are busy. Increase ``pm.max_children`` if RAM
+  allows. Enable the slow log to check whether a slow query is tying up workers.
+  Also check that a ``request_terminate_timeout`` is not killing workers mid-request.
+
+**504 Gateway Timeout**
+  A worker is running but not responding within the web server's upstream timeout
+  (nginx ``fastcgi_read_timeout``, Apache ``ProxyTimeout``). Common Nextcloud causes:
+  large file operations, slow database queries during sync, or PROPFIND over large
+  directory trees. Use the slow log to identify the bottleneck.
+
+**Memory grows over time**
+  Memory leaks can occur in worker processes. Common culprits include libraries that manage external resources (like Imagick for image processing). Use the slow log and RSS monitoring to identify which requests cause growth.
+  Set ``pm.max_requests = 500`` to recycle them before they grow too large.
+
+**Slow first request after idle**
+  ``pm = ondemand`` or ``pm.min_spare_servers`` too low. Switch to ``pm = dynamic``
+  and raise ``pm.min_spare_servers``.
+
+After any configuration change, reload PHP-FPM — changes do not take effect until you do:
+
+.. code-block:: bash
+
+   sudo systemctl reload php8.3-fpm   # Debian/Ubuntu — adjust version as needed
+   sudo systemctl reload php-fpm      # RHEL/Fedora
+
+For pool configuration details (environment variables, upload sizes, Unix socket vs TCP),
+see :ref:`php_fpm_tips_label` in the installation guide.
 
 Enable PHP OPcache
 ------------------